Grammatical Facial Expressions Recognition

Train the classifier on expressions from user A.
Use an appropriate classification method among those you know (e.g. hard SVM, soft SVM, K-NN, CNN,
etc.) to train the classifier on expressions from user A, broken down into:
a. Basic implementation, using an off-the-shelf classifier (e.g. a standard R command), on a
single facial expression.
b. Repeat the training for a different facial expression.
c. Extra marks for coding your own implementation of the chosen classifier.
2. Evaluate accuracy.
Employ this trained classifier on the expressions performed by user B, and determine the accuracy,
broken down into:
a. Test on a single facial expression, using an off-the-shelf classifier, and comment on the results.
b. Test on a different facial expression, and comment on the difference.
c. Use of performance measures other than simple accuracy (e.g. precision, recall, Fscore, ROC
curve)
3. Additional experiments on the first classification method
a. Repeat the analysis (training + testing) by inverting the roles of the user (i.e., training on user
B and testing on user A), and comment on the difference
b. Repeat the analysis (training + testing) by considering a different feature representation than
the original landmark coordinate vector, and comment. For instance, you can use the
two representations proposed in the paper by Freitas et al, you can try PCA or any other
dimensionality reduction, and so on.
4. Second ML Technique and Performance Evaluation
Choose a different classifier than before, and repeat the training and testing, broken down into:
a. Basic training of an off-the-shelf classifier on a single facial expression
b. Testing of an off-the-shelf classifier on a single facial expression, commenting on the
results;
c. Extra marks for own implementation of the second classifier
d. Training the second classifier on an extra expression
e. Testing the second classifier on an extra expression, with comments
f. Repeat the analysis (training + testing) by inverting the roles of the user (i.e., training on user
B and testing on user A), and comment on the results
g. Repeat the analysis (training + testing) by considering a different feature representation than
the original landmark coordinate vector, and comment on the results
h. Use of performance measures other than simple accuracy
5. Conclusion
Compare the results of the two classifiers, and comment on their performance. Summarise your main
findings. Explain possible current limitations of your solutions and possible further strategies to
improve on the results. All arguments must be evidence-based.
THE WORK NEEDS TO BE DONE IN (R)